Reading Recognition Method of Mechanical Water Meter Based on Convolutional Neural Network in Natural Scenes

Abstract

To satisfy the demand for real-time and high-precision recognition of mechanical water meter readings in natural scenes, a reading recognition method for mechanical water meters based on you only look once version 4 (YOLOv4) is proposed in this paper. First, a focus structure is introduced into the feature extraction network to expand the receptive field and reduce the loss of original information. Second, a ghost block cross stage partial module is constructed to improve the feature fusion of the network and enhance the feature representation. Finally, the loss function of YOLOv4 is improved to further enhance the detection accuracy of the network. Experimental results show that the mAP@0.5 and mAP@0.5:.95 of the proposed method are 97.9% and 77.3%, respectively, which are 1.6% and 6.0% higher, respectively, than those of YOLOv4. Additionally, the number of parameters and computation amount of the proposed method are 48.6% and 36.8% lower, respectively, whereas its inference speed is 27% higher. The proposed method is applied to assist meter reading, which significantly reduces the workload of on-site meter-reading personnel and improves work efficiency. The datasets used are available at https://github.com/914284382/Mechanical-water-meter.